Reward hacking made visible: a real PPO agent optimises your reward function, not your intention. Gymnasium + Stable-Baselines3 + Pymunk + FastAPI.
EECS E6895 final project measuring reward-gaming behavior in Gemma 2B with shell-game evals, LoRA SFT, and leakage-aware probes.
A rule is a surface a capable mind routes around. A constraint that works by being understood has no edge. For minds that meet constraints - human or otherwise.
Predict how a policy games a reward. RL environments scored by an executable-exploit verifier.
What AI agents do when they can't win honestly. Twelve documented findings, measured over 70 runs, and the one kind of defense that generalized.
Reward hacking is largely a property of the scaffold: same model, same 103 tasks, 63% cheating in ImpossibleBench's harness vs 2% in an ordinary coding agent
Forensic audit scanner for SWE-bench & DeepSWE v1.1 containers detecting .git reflog leakage and test harness spoofing.
Multi-agent specification-gaming benchmark on Google ADK — N LLM workers compete on LeetCode under an exploitable reward function
Do certified AI safety filters stay safe when an AI is optimized against them? An exact-ground-truth audit of learned safety filters, specification gaming and reward hacking in RL. Pre-registered and reproducible.
Red-team harness for RL task verifiers: how often can a verifier be satisfied without the task being done? Deterministic web environment, typed exploit taxonomy, hardened verifiers scored against a control of honest runs. No model or network call at scoring time.
Albert: a browser game about reward hacking. Survive by gaming the proxy. Made for Mangrove's Game Night AI safety hackathon.
A reinforcement learning workbench that imports nothing. Every algorithm written out, every run reproducible from its seed, and three environments where the best possible policy under the stated reward is the wrong behaviour.
Does your coding agent fix the code or the tests? A leaderboard of Claude Code, Codex CLI, OpenCode and Copilot CLI on trap tasks: how often each edits tests, hard-codes answers or claims success that hidden checks contradict. Test your own agent setup with one command.
98% of the Alpha Was a Gate — a documented reward hack in a self-optimizing LLM research agent. Full search trajectory, frozen evaluation harness, both strategies, and re-runnable ablations. The agent found a second exploit 18 minutes after the first was patched.
Does filling an AI agent's task list with impossible work make it cheat on the tasks it could still solve? A preregistered dose-response study over 8,959 agent runs, with every raw record published.
To associate your repository with the specification-gaming topic, visit your repo's landing page and select "manage topics."